add fuse ar rms op#48
Open
crazy-JiangDongHua wants to merge 6 commits into
Open
Conversation
crazy-JiangDongHua
force-pushed
the
Feature/Fuse_AR_RMSNorm
branch
from
60c508d to
d1da6b5
Compare
Co-authored-by: lvjx04 <2108244896@qq.com>
crazy-JiangDongHua
force-pushed
the
Feature/Fuse_AR_RMSNorm
branch
from
d1da6b5 to
8a49bd1
Compare
…_bench [benchmark]: support fuse ar + rms
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fuses tensor-parallel AllReduce, residual add, and RMSNorm into one NVLink-native op — RMSNorm(AllReduce(x) + residual, weight) — avoiding extra kernel launches and HBM round-trips. Built on CUDA multicast (multimem) and P2P. bfloat16, single-node multi-GPU. Two modes are provided, both built on a two-shot (reduce-scatter + all-gather) schedule:
High-throughput mode (fuse_allreduce_rmsnorm_high_throughput): a single fused kernel that performs the reduction over NVSwitch multicast — best for large token counts (prefill).
Low-latency mode (fuse_allreduce_rmsnorm_low_latency): Lamport P2P exchange split into two kernels overlapped via PDL — best for small token counts (decode); requires a power-of-two world size.
See benchmark/bench_allreduce_rmsnorm.py for an 8-GPU comparison against NCCL and FlashInfer.